callspunkbuster

GameBotDetectionBasedonAvatarTrajectoryKuan-TaChen1,AndrewLiao2,Hsing-KuoKennethPao3,andHao-HuaChu21InstituteofInformationScience,AcademiaSinica2Dept.
ofComputerScience&InformationEngineering,NationalTaiwanUniversity3Dept.
ofComputerScience&InformationEngineering,NationalTaiwanUniv.
ofScience&TechnologyAbstract.
Inrecentyears,onlinegaminghasbecomeoneofthemostpopularInternetactivities,butcheatingactivity,suchastheuseofgamebots,hasincreasedasaconsequence.
Generally,thegamingcommunitydisagreeswiththeuseofgamebots,asbotusersobtainunreasonablerewardswithoutcorrespondingeorts.
However,botsarehardtodetectbecausetheyaredesignedtosimulatehumangameplayingbehaviorandtheyfollowgamerulesexactly.
Existingdetectionapproacheseitherinterrupttheplayers'gamingexperience,ortheyassumegamebotsarerunasstandaloneclientsorassignedaspecicgoal,suchasaimbotsinFPSgames.
Inthispaper,weproposeatrajectory-basedapproachtodetectgamebots.
Itisageneraltechniquethatcanbeappliedtoanygameinwhichtheavatar'smovementiscontrolleddirectlybytheplayers.
Throughreal-lifedatatraces,weshowthatthetrajectoriesofhumanplayersandthoseofgamebotsareverydierent.
Inaddition,althoughgamebotsmayendeavortosimulateplayers'decisions,certainhumanbehaviorpat-ternsarediculttomimicbecausetheyareAI-hard.
TakingQuake2asacasestudy,weevaluateourscheme'sperformancebasedonreal-lifetraces.
Theresultsshowthattheschemecanachieveadetectionaccu-racyof95%orhighergivenatraceof200secondsorlonger.
Keywords:CheatingDetection,OnlineGames,Quake,Security,Su-pervisedClassication,UserBehavior.
1IntroductionInrecentyears,onlinegaminghasbecomeoneofthemostpopularInternetactiv-ities.
However,asthepopulationofonlinegamershasincreased,gamecheatingproblems,suchastheuseofgamebots,havebecomemoreserious.
GamebotsThisworkwassupportedinpartbyTaiwanInformationSecurityCenter(TWISC),NationalScienceCouncilunderthegrantsNSC97-2219-E-001-001andNSC97-2219-E-011-006.
ItwasalsosupportedinpartbyTaiwanE-Learning&DigitalArchivesProgram(TELDAP),NationalScienceCouncilunderthegrantsNSC96-3113-H-001-010andNSC96-3113-H-001-012.
S.
M.
StevensandS.
Saldamarco(Eds.
)ICEC2008,LNCS5309,pp.
94–105,2008.
cIFIPInternationalFederationforInformationProcessing2008GameBotDetectionBasedonAvatarTrajectory95areautomatedprogramswitharticialintelligencethatplayersusefordierentpurposes.
InMMORPGs(MassivelyMultiplayerOnlineRolePlayerGames),playerscansaveagreatdealoftimebyusingbotstoperformrepetitivetasks,suchasslashinglow-levelmonsters,orshinginarivertomastertheavatar'sshingskills.
InFPS(First-PersonShooter)games,userscanemploybotstoplayinplaceofthemselvesinordertogethighscoresandgainareputationinthecommunity.
Generally,thegamingcommunitydisagreeswiththeuseofgamebots,asbotusersobtainunreasonablerewardswithoutcorrespondingeorts.
However,gamebotsarehardtodetectbecausetheyaredesignedtosimulatehumangameplayingbehaviorandtheyfollowgamerulesexactly.
Somebotdetectionstud-ies[1,2]proposeusingCAPTCHAtestsduringagametodeterminewhetheranavatarisactuallycontrolledbyaperson.
Althoughthismethodiseective,itinterruptsthegameplayanddegradesplayers'feelingsofimmersioninthevirtualworld[3,4].
Alternatively,passivedetectionapproaches,suchasschemesbasedontracanalysis[5,6]andschemesbasedonavatars'shootingaccuracyinFPSgames[7],canbeused.
Theformerapproachassumesthatagamebotworksasastandaloneclient,andthelatterisonlyvalidfordetectingaimbotsinshootinggames.
Inthispaper,weproposeageneralapproachforallgenresofgameswhereplayerscontroltheavatar'smovementdirectly.
Ourapproachisbasedontheavatar'smovementtrajectoryduringagame.
Therationaleisthatthetrajectoryoftheavatarcontrolledbyahumanplayerishardtosimulate.
Playerscontrolthemovementofavatarsbasedontheirknowledge,experience,intuition,andagreatdealofinformationprovidedinthegame.
Sincehumandecisionsmaynotalwaysbelogicalandecient,howtomodelandsimulaterealisticmovementsisstillanopenquestionintheAIeld.
Todistinguishhumanplayersfromgamebotseciently,weanalyzethetrajectoriesofbothplayertypesanddistinguishbetweenthetrajectoriesaccordingtotheirspatialandtemporalcharacteristics.
WechooseQuake2asourcasestudybecauseitisaclassicandpopularFPSgame,andmanyreal-lifehumantracesareavailableontheInternet.
Therefore,wecanusesuchtracestovalidateourproposedscheme.
Thecontributionofthispaperistwo-fold.
1)Weproposeatrajectory-basedapproachfordetectinggamebots.
Itisageneralmodelthatcanbeappliedtoanygameinwhichtheavatar'smovementiscontrolledbytheplayersdirectly.
2)Usingreal-lifehumantraces,theperformanceevaluationresultsshowthattheschemecanachieveadetectionaccuracyof95%orhigherwhenthetracelengthis200secondsorlonger.
Becauseitisdiculttosimulatehumanplayers'logicwhentheycontrolgamecharacters,webelievethisapproachhasthepotentialtodistinguishbetweenhumanplayersandautomatedprogramsandthusmeritsfurtherinvestigation.
Theremainderofthispaperisorganizedasfollows.
Section2containsareviewofrelatedworks.
InSection3,weintroduceourgamecasestudy,Quake2,anddescribethegametracecollectionmethodology.
WeanalyzethesimilaritiesanddierencesbetweenthetrajectoriesofdierenttypesofplayersinSection4.
In96K.
-T.
Chenetal.
Section5,weproposeanidenticationschemeanddemonstrateitsabilityintermsofthedistributionofdiscriminativefeatures.
InSection6,weevaluatetheperformanceoftheproposedschemewiththeconsiderationofthetracelength.
Then,inSection7,wesummarizeourconclusions.
2RelatedWorkRecently,anti-cheatingsoftwareprograms,suchasPunkBusterandGameGuard,havebeenwidelydeployedinonlinegamestopreventcheating.
Suchsoftwareisbundledwithgameclients,soitcannotbeuninstalledevenifthegameclienthasbeenremoved.
Itworksbyhidinginthegameclientprocess,monitoringtheentirevirtualmemoryspace(topreventmodicationofthegame'sexecutableimages),blockingsuspectedprogramsthatmightbehackertools,andblockingcertainAPIcalls.
Thiskindofsoftwarecandetectnearlyallplug-intoolsthatattachtoagameclientprogramtoinspectormodifygamestateswhenthegameisrunning.
Unfortunately,itcannotstopthewidespreaduseofstandalonebots,includingthebotserieswestudyinthispaper.
Thereasonisthattheseanti-cheatingsoftwareprogramsarehost-based,sotheymustbeinstalledonplayers'PCstobeeective.
Standalonebots,ontheotherhand,canfunctionwithoutclients,anditisunlikelythatanti-cheatingtoolswouldbeinstalledonPCswherethebotsarerunning.
ThisclaimisstronglysupportedbythefactthatgamebotsarestillactiveingamesprotectedbyPunkBusterorGameGuard,e.
g.
,Quake(PunkBuster)andLineage1(GameGuard).
3DataDescription3.
1HumanTracesQuake2supportsagame-playrecordingfunctionthatkeepstrackofeveryactionandmovement,aswellasthestatusofeachcharacteranditemthroughoutthegame.
Witharecordedtrace,onecanreconstructagameandreviewitfromanypositionandangledesiredviaVCR-likeoperations.
Playersoftenusethisfunc-tiontoassesstheirperformanceandcombatstrategies.
Moreover,experiencedplayersareencouragedtopublishtheirgame-playtracesasteachingmaterialsfornovicegamersandtherebybuildareputationinthecommunity.
ToensurethatourgametracesrepresentedthediversityofQuakeplayers,weonlyusedtracesthatplayershadcontributedvoluntarily.
Thehumanplayers'tracesweredownloadedfromthefollowingarchivesites:GotFragQuake2,PlanetQuake3,DemoSquad4,andRevillaQuakeSite5.
WerestrictedthetracestothemapTheEdge,oneofthemostwell-knownlevelsofdeath-matchplay.
On1http://boards.
lineage2.
com/showflat.
phpNumber=5737372http://www.
gotfrag.
com/quake/home/3http://planetquake.
gamespy.
com/4http://q2scene.
net/ds/5http://www.
revilla.
nildram.
co.
uk/demos-full.
htmGameBotDetectionBasedonAvatarTrajectory97Table1.
TraceSummarynamenummeantotalactive1Human932hour203.
5hour91%2CR2419hour448.
8hour91%3Eraser1520hour296.
4hour94%4ICE1820hour358.
8hour67%thismap,theonlygoalisthateachplayershouldkillasmanyotherplayersaspossible,untilthetimelimitisreached.
Becauseshorttracescontainlittleinformation,weonlycollectedtraceslongerthan600seconds.
3.
2BotTracesTherearemanygamebotsavailableforQuake2.
Forthisstudy,weselectedthreeofthemostpopularbotprogramsfortracecollection,namelyCRBOT1.
14[8],EraserBot1.
01[9],ICEBot1.
0[10].
Wecollected1,306hoursoftracesintotal,asshowninTable1.
InCRBotandEraserBot,allhumanplayersandbotswereactivemostoftime(≥90%).
TherewaslessactivityinICEBotbecauseitoftenremainedidleinsomeplaceswaitingforanopportunitytoambushotherplayers.
4DiscriminativeAnalysisInthissection,wecomparetheavatartrajectoriesofhumanplayersandgamebots.
First,wecomparethenavigationpatternsofthetwoplayertypesandconsidertheirindividualtrajectories.
Wethenidentifythemostsignicantdis-criminativecharacteristicsoftherespectivetrajectoriesandincorporatethemintotheproposedbotidenticationscheme.
Weconstructtheaggregatednavigationpatternofeachplayertypebyplottingalltheobservedcoordinatesinalltracesoftheparticularplayertypeonagraph,asshowninFig.
1.
Theareasofhighdensityineachgurearetheplacesthatplayersvisitmorefrequently,whilethesparseareasrepresentbuildingsorothertypesofobstaclesthatplayerscannotpass.
Theguresshowthatthegamelevelisformedbysquares,plazas,andnarrowcorridors.
Thisarrangementisdesignedspecicallyfordeath-matchplay,asthewindingroutesprovidecoverforplayerstohide,andthenarrowcorridorsleadtointenseghtingifplayersconfronteachotherintheseplaces.
Weobservethat,eventhoughallthemovementtraceswerecollectedonthesamemap,thenavigationpatternsofdierentplayertypesaredissimilar.
Wesummarizethedierencesbelow.
1.
Humanplayerstendedtoexploreallareasonthemap;thus,Fig.
1(a)showsthemostcompleteterrainofthelevel.
Incontrast,theroutingalgorithmsofgamebotsmayhavehaddicultynavigatingtosomeplaces,sotheynevervisitedsomepartsofthemap.
Forexample,thebottomleft-handcorneroftheCRBotnavigationmapinFig.
1(b)doesnotindicatethepresenceofbots.
98K.
-T.
Chenetal.
(a)Humanplayers(b)CRBot(c)EraserBot(d)ICEBotFig.
1.
Presencelocationsofallplayers2.
Toreducetheprobabilityofbeingattacked,humanplayersnormallyavoidopenspaces.
Therefore,inFig.
1(a)weobservethathumanplayersavoidedtheplazainthemiddleofthemap,andstayedinthesurroundingcorridorsinstead.
Thisisindicatedbythehighdensityofplotsinthecorridors.
Incontrast,gamebotsoftenstayinthecentralplaza,probablybecauseitoccupiesalargespaceanditiseasytogeteverywherefromthisarea.
3.
Eventhoughhumanplayersspendmostoftheirtimeinnarrowareasandconnedrooms,therearelargevariationsintheirtrajectories.
Therearetworeasonsforthisphenomenon.
1)Thewidthofthemainroutesisquitelarge.
Ratherthanstayinthemiddleofaroute,playersmoveirregularlywithinthelimitedspace.
Thismaybeduetoplayers'preferences;hence,someplayersmaymovealongthewallofthepath,whileothersmaywalkstraight,unlesstheavatarisblockedbyawallorotherobstacles.
2)Asghtsmayoccuranytime-anywhere,humanplayersoftenmovestrategicallytododgecurrentorpotentialattacks.
Ontheotherhand,wendthatdierentgamebotsadoptverydierentmovementpatternsovertheroutes.
ThemovementpathsofCRBotandEraserBot(Fig.
1(b)andFig.
1(c)respectively)aredenseandeasytosee.
Thissuggeststhatthesebotstendtofollowexactmovementpatternswhenmovingthroughthesamecorridor.
Incontrast,ICEBot(Fig.
1(d))exhibitsanearlyuniformdistributionoverallpossiblepointsonthemap.
Thisimpliesthatitsroutingalgorithmdecidestheavatar'sGameBotDetectionBasedonAvatarTrajectory99directionratherthanitsexactmovementpattern,sothattheprobabilitiesofallpointsontherouteareroughlyequivalent.
Clearly,therearesubstantialdierencesbetweentheaggregatednavigationpatternsofhumanplayersandthoseofeachgamebotbecausethebots'rout-ingpatternsareverydierentfromthemovementbehaviorexhibitedbyhumanplayers.
5BotDetectionSchemeOurobjectiveistoclassifyhumanplayersandgamebotsecientlyandac-curately.
Tothisend,weintegratethespatialandtemporaldierencesinthetrajectoriesofavatarscontrolledbydierentplayertypestodevelopabotidenti-cationscheme.
Inthissection,werstdescribethesetofdiscriminativefeaturesextractedfromtheavatartrajectories,andthenexplainhowweusethefeaturestoclassifygamebotsandhumanplayers.
5.
1FeatureExtractionGivenasegmentofatrajectory,{xt,yt},1≤t≤T,weextractthefollowingfeaturesfromthistwo-dimensionaltimeseries.
1.
ON/OFFActivity.
First,wenotethatavatarsinthegameplaydonotmoveallthetime.
Sometimestheymaystoptocheckifanyopponentsarearound,waitforopponentstoenteranarea,waitforregenerationoftheirweaponsorammunition,orsimplytakearest.
ThealternatemovingandidlebehaviorformsanON/OFFmovementpattern.
WedeneONperiodsasconsecutiveperiodsofmovementlongerthan1second,andOFFperiodsastheremainingtimeframes.
ThedurationandfrequencyofON/OFFperiodsaredecidedbytheplayers'stylesandthebots'AIlogic.
Forexample,aggressiveplayersmaykeepmovingallthetime,whilecautiousplayersmaystayinoneplacetomonitortheirsurroundings.
Therefore,wedenefourfeaturesbasedonON/OFFactivity:themeanandstandarddeviationofONperiods,andthoseofOFFperiods.
Fig.
2showsthedistributionsofthefourfeatures.
Themeanandstandarddeviationofhumanplayers'ONperiodsaresignicantlyhigherthanthoseofgamebots.
Thisindicatesthathumanplayersaremoreaggressiveastheytendtomoveallthetime.
Inaddition,themeanandstandarddeviationofhumanplayers'OFFperiodsarelongerthanthoseofbots,whichimpliesthathumanbehaviorismoreirregularandunpredictableinthattheymaywaitforalongertimeafteralongmove.
ThegureshowsthathumanplayersandgamebotsdierintermsofON/OFFactivity.
Hence,webelievethatthefourfeaturesbasedontheseactivitiescouldbeusefulforbotdetection.
2.
Pace.
Ingames,avatarsaregenerallyallowedtomoveatdierentspeedsandindierentways,suchasrunning,slowwalking,step-by-stepwalking,lateral100K.
-T.
Chenetal.
HumanCREraserICE050150250OnPeriodMeanHumanCREraserICE050100200OnPeriodSDHumanCREraserICE5101520OffPeriodMeanHumanCREraserICE05152535OffPeriodSDFig.
2.
ThedistributionoffeaturesrelatedtoON/OFFperiodsshifting,andmovingbackwards.
Inaddition,playerscanstopthecurrentmove-mentandproceedwithanothermovementindierentdirectioninsub-seconds;therefore,theresultingavatarmovementscanbehighlyvariable.
Onesimplewaytocharacterizethedynamicsofanavatar'smovementisbythepaceofitsmovements.
Wedenethepaceasthedisplacementofanavatar'scoordinateinonesecond,andextractthemeanandstandarddeviationofthepaceastwofea-tures.
Wendthatthepacesofmostavatarsaregenerallysmall,althoughtheycanbelargeoccasionally.
Tocharacterizethevariabilityofpaceswhenplayersmovefast,wealsodenethe"largepaceSD,"whichisthestandarddeviationofpaceslargerthan10units.
Inadditiontonormalmovements,playersmayteleporttheiravatarstoaremoteplaceinstantlythroughateleportationspot.
Teleportationmayalsobeusedwhenanavatardies.
Itwillbetransferredtotherebirthspotsothatitslifepointscanberecharged.
Wedetectteleportationoccurrencesbycomputingiftheosetinonesecondislongerthan60unitsanddenethefeature"teleportationrate"astheaveragecountofteleportationoccurrencesrecordedinonesecond.
Fig.
3showsthedistributionofthefourfeaturesrelatedtothemovementpaceandteleportation.
Althoughthemeansofthepacesofdierentplayertypesaredissimilar,thevariationsarenotlarge.
Thisshowsthatthefourplayertypeshavedierentbutconsistentmicro-movementbehaviorinsmalltimescales.
Thestandarddeviationofthepacealsohaslargediscriminability,wherethatofhumanplayersandEraserBothavesimilarmagnitude.
Thelargestandarddeviationofthepace,ontheotherhand,exhibitsgreatdiscriminability,whichindicatesthathumanplayershaveevenlargerpacevariabilitywhentheymovefast.
Finally,CRBotandEraserBothaveverylowteleportationfrequency.
Incontrast,humanplayershavemoderateteleportationfrequency.
Moreover,theirGameBotDetectionBasedonAvatarTrajectory101HumanCREraserICE51015202530PaceMeanHumanCREraserICE24681012PaceSDHumanCREraserICE246810Pace(>10)SDHumanCREraserICE0.
000.
040.
08TeleportationFig.
3.
Thedistributionoffeaturesrelatedtomovementpacevarianceishighbecausehumanplayershavedierentpreferenceswhenusingteleportationspotsandplayersgetkilledatdierentrates.
.
3.
PathWealsodenethefollowingfeaturestocharacterizethedetailedtra-jectoriesofavatarsinagame.
Lingering.
Weconsiderwhetherplayers"lingered"inasmallareaduringaspecictimeperiod.
Foranavatarat(x,y)attimet,ifitsdistancefrom(x,y)wasalwayslessthandduringtheperiod(t,t+p),wesaythattheavatarwaslingeringduring(t,t+p),giventheparameters(d,p).
Wearbitrarilysetd=30secondsandp=300units,aswendthatdierentparametersdonotaecttheclassicationperformancesignicantly.
Smoothness.
The"smoothness"featuredetermineswhetheranavatarmovesinstraightorzig-zagpatterns.
Assumeanavatarisat(x1,y1)attimet1andat(x2,y2)attimet2.
Wedenethesmoothnessasthenumberoftimesthecharactermovesacrosstheline(x1,y1)(x2,y2)duringtheperiod(t1,t2).
Astheline(x1,y1)(x2,y2)indicatestheshortestroutebetweenthetwopoints(x1,y1)and(x2,y2),crossingthelineimpliesthattheplayerismovingineciently.
Thismaybebecauseheisattemptingtododgegunre,switchtoanothertarget,orsimplyduetoplayers'habitsorbots'routingalgorithms.
Detour.
Wedeneanotherfeature"detour"toquantifytheeectivenessofusermovements.
Ifanavatarisat(x1,y1)attimet1andat(x2,y2)attimet2,wecomputethedetourbydividingthelengthofthemovementbytheeectiveosetofanavatarduringtheperiod(t1,t2).
ThedistributionsoftheabovefeaturesareplottedinFig.
4.
Thegraphshowsthatthelingerfrequencyanddurationofhumanplayersaresignicantlyless102K.
-T.
Chenetal.
HumanCREraserICE0.
010.
030.
05LingerFrequencyHumanCREraserICE1520253035LingerLengthHumanCREraserICE0.
800.
901.
00SmoothnessHumanCREraserICE51015DetournessFig.
4.
Thedistributionoffeaturesrelatedtomovementpaththanthoseofgamebots.
Thisisreasonablebecauselingeringinaplaceforalongtimeisadangerous,astheplayermaybenoticedandinduceopponents're.
Thesmoothnessofhumanplayersisthelowestofthefourplayertypes,whichsupportstheintuitionthathumanplayers'movementsarethemostirregularandunpredictable.
ThedetourfeatureshowsthatEraserBotmovesveryinecientlyintermsoftheavatar'seectiveoset.
Incontrast,themovementsofhumanplayersarerelativelymoreecient.
Wesuspectthisisbecausehumanplayerstendtomoveawayfromcurrentpositionstoanotherplaceecientlyeventhoughtheymaymoveirregularlyandstrategically;thus,theresultingavatartrajectoryexhibitsbothunpredictabilityandeciencywhichseemcontradictory.
4.
Turn.
Ournalsetoffeaturesisbasedonthefrequencyandamplitudeofhowavatarschangedirection.
Ourrationaleisthateachtimeanavatarchangesdirection,themagnitudeofthechangeshouldbedependentonplayerconven-tionsandbotroutingalgorithms.
Assumeanavatarisat(x1,y1)attimet,at(x2,y2)attimet+p,andat(x3,y3)attimet+2p.
Iftheanglebetweentwovectors(x2x1,y2y1)and(x3x1,y3y1)isgreaterthana,wedeterminethataturnwithangleaoccurred.
Wedenethreefeaturestodenotethefrequencyofturnswithangles30,60,and90,respectively.
Inaddition,wedeneafeaturecalledthe"turnangle"todenotetheaverageanglechangeforalldirectionchangesgreaterthan30.
Fig.
5showsthedistributionsoftheturn-relatedfeatures.
Weobservethatthefourplayertypeschangedirectionatdierentratesnomatterhowwedenetheminimumdegreeofadirectionchange.
Notably,theturnfrequencyofhumanplayersisthehighestforthe30angleandbecomesrelativelylowerforthe90angle.
Inaddition,theaverageturnangleofhumanplayersisthelowestamongGameBotDetectionBasedonAvatarTrajectory103HumanCREraserICE0.
00.
20.
40.
6Turn30HumanCREraserICE0.
00.
20.
40.
6Turn60HumanCREraserICE0.
00.
20.
4Turn90HumanCREraserICE60708090110TurnAngleFig.
5.
Thedistributionoffeaturesrelatedtoturnmovementthefourtypes,whichindicatesthathumanplayerstendtoadjusttheirdirectionscontinuouslyandslightly.
5.
2ClassicationWeapplyasupervisedclassicationframeworktotrainaclassier,whichweusetodeterminewhetherasegmentofanavatar'strajectorybelongstoahumanplayeroragamebot.
TheclassierweadoptisthenaiveBayesianclassierwithoutthekerneldensityestimation.
Weevaluatetheperformanceoftrajectoryclassicationinthenextsection.
6PerformanceEvaluationInthissection,weevaluatetheperformanceofourproposedbotdetectionschemeonthecollectedtraces.
First,weevaluatewhetherourschemecandistin-guishbetweenhumanplayersandgamebots,byusingtheclassiertoperform10-foldcross-validation.
Inreal-lifescenarios,thetracelengthplaysanimpor-tantrolebecauseitdetermineshowquicklyagamebotcanbedetected.
Thus,weevaluatedtheperformanceofourschemeondierenttraceslengths,asshowninFig.
6.
Thegraphshowsthatthedetectionaccuracyishigherthan90%,evenwhenthetracelengthisasshortas100seconds.
Longertracesyieldbetteraccuracy.
Todeterminewhichcategoryoffeaturesyieldsthehighestaccuracy,weplottheclassicationperformanceforeachcategoryoffeatures.
Theresultsindicatethatthefeaturesrelatedtothemovementpace,directionchanges,andON/OFFperiodsallyieldgoodresults,whilepath-relatedfeaturesonlyexhibitgooddiscriminabilitywhenthetracelengthis800secondsorlonger.
104K.
-T.
Chenetal.
20040060080010000.
00.
20.
40.
60.
81.
0Observationtime(sec)Accuracy1002003004005006007008009001000ON/OFFfeaturesPacefeaturesPathfeaturesTurnfeaturesAllfeaturesFig.
6.
Accuracybetweenhumanandbots20040060080010000.
00.
20.
40.
60.
81.
0Observationtime(sec)Accuracy1002003004005006007008009001000ON/OFFfeaturesPacefeaturesPathfeaturesTurnfeaturesAllfeaturesFig.
7.
Classicationaccuracybetweenfourtypesofplayers(humanandthreebotprograms)Furthermore,weperformaplayer-typeclassication;thatis,wenotonlydeterminewhetheracharacteriscontrolledbyahumanplayerorabotprogram,butalsoidentifythebotprogramusedifappropriate.
TheresultsareshowninFig.
7.
Theclassicationaccuracyoftheplayertypesisevenbetterthanthatofthehuman-botscenariowhenthetracelengthislongerthan200seconds.
Withatracelengthof500secondsorlonger,ourschemeyieldsaclassicationaccuracyof98%orhigher.
However,inthissetting,individualfeaturecategories,exceptthoserelatedtomovementpaces,exhibitlowdiscriminabilitywhentheyareappliedtotheclassicationseparately.
7ConclusionWehaveproposedatrajectory-basedapproachfordetectinggamebots.
Itisageneraltechniquethatcanbeappliedtoanygameinwhichtheavatar'sGameBotDetectionBasedonAvatarTrajectory105movementiscontrolledbytheplayersdirectly.
Ouranalysisofreal-lifetracesshowsthatthetrajectoriesofhumanplayersandgamebotsareverydissim-ilar.
Theperformanceevaluationresultsshowthatourbotdetectionschemecanachieveadetectionaccuracyof95%orhigherwhenthetracelengthis200secondsorlonger.
Becauseitisdiculttosimulatehumanplayers'behaviorwhencontrollinggamecharacters,webelieveourmethodhasthepotentialtodistinguishbetweenhumanplayersandautomatedprograms,andthusmeritsfurtherinvestigation.
References1.
Golle,P.
,Ducheneaut,N.
:Preventingbotsfromplayingonlinegames.
ComputersinEntertainment3(3),3(2005)2.
vonAhn,L.
,Blum,M.
,Hopper,N.
J.
,Langford,J.
:CAPTCHA:UsinghardAIproblemsforsecurity.
In:ProceedingsofEurocrypt,pp.
294–311(2003)3.
Novak,T.
P.
,Homan,D.
L.
,Duhachek,A.
:Theinuenceofgoal-directedandexperientialactivitiesononlineowexperiences.
JournalofConsumerPsychol-ogy13(1),3–16(2003)4.
Ila,S.
,Mizerski,D.
,Lam,D.
:Comparingtheeectofhabitintheonlinegameplayofaustralianandindonesiangamers.
In:ProceedingsoftheAustraliaandNewZealandMarketingAssociationConference(2003)5.
Chen,K.
T.
,Jiang,J.
W.
,Huang,P.
,Chu,H.
H.
,Lei,C.
L.
,Chen,W.
C.
:IdentifyingMMORPGbots:Atracanalysisapproach.
In:ProceedingsofACMSIGCHIACE2006,LosAngeles,USA(June2006)6.
Chen,K.
T.
,Huang,P.
,Lei,C.
L.
:Gametracanalysis:AnMMORPGperspective.
ComputerNetworks50(16),3002–3023(2006)7.
Yeung,S.
,Lui,J.
,Liu,J.
,Yan,J.
:Detectingcheatersformultiplayergames:theory,designandimplementation.
ProcIEEECCNC6,1178–11828.
Malakhov,M.
:CRBot1.
15(May2000),http://arton.
cunst.
net/quake/crbot/9.
Feltrin,R.
R.
:EraserBot1.
01(May2000),http://downloads.
gamezone.
com/demos/d9862.
htm10.
jibe:ICEBot1.
0(1998),http://ice.
planetquake.
gamespy.
com/